Key Technologies Of Face Attribute Analysis

Face attributes are the specific description of facial features.In this thesis,the algorithm of face attribute recognition and the compression method of face attribute recognition model are studied deeply,and a demonstration system of face attribute recognition is implemented.The main achievements of the thesis are:1.A method of grouping attributes based on attribute relationship is proposed,so that the attributes in one attribute group can learn from each other.In the meantime,in order to extrate more discriminative attribute features,two constraints are designed for the ouput of the attribute model,namely,intra-attribute ranking and inter-attribute ranking.Experimental results show that the proposed method can make the network easier to train while improving the accuracy of face attribute recognition;2.Scratch-based method for model compression,knowledge distillation algorithm and automated quantization algorithm.A lightweight network structure for face attribute recognition is proposed,and a large attribute model with high precision is used to supervise the optimization of parameters in the small network structure,so that the small model can achieve similar recognition performance of the large model.The experimental results show that a large amount of computation cost in the forward running of the model can be reduced at the loss of a small part of the attribute recognition accuracy;3.A face attribute recognition system is implemented.The main functions of the system are face detection and face attribute recognition.The system can run in real time and accurately recognize all face attributes in the video streams and images.